SpecForge

Project Name SpecForge Summary SpecForge is an evidence-grounded multi-agent council built to solve a real 3GPP pain: when teams interpret complex specs, a single AI answer can miss conditions, invent assumptions, or produce implementation guidance that is unsafe to ship. SpecForge uses multiple specialist agents, evidence retrieval, and a judge workflow to produce more reliable, decision-grade outputs. Project Description Our starting point was a real 3GPP standards problem. Engineers working with telecom specs often deal with dense requirements, exceptions, cross-references, and vendor-specific ambiguity. In that environment, a single LLM response is not enough. It may sound confident, but it can still miss mandatory headers, mix models, flatten edge cases, or hide uncertainty. SpecForge fixes that by turning AI into a council instead of a single answer engine. It brings together multiple specialist agents to debate the question, ground claims in evidence, and generate a final synthesis that is far more trustworthy. It also benchmarks Single LLM, Single LLM + Evidence, and Full Council + Jury side by side, then visually shows where the weaker approaches made mistakes and how the council corrected them. While it was born from 3GPP and telecom standards pain, the same system also extends naturally to code review, RCA, sprint planning, and other high-stakes decision workflows. Key Features * Built around real 3GPP standards interpretation pain * Multi-agent council with evidence-grounded debate * Side-by-side comparison of Single LLM vs Council workflows * Judge Board showing mistakes, contradicted assumptions, and corrections * Rich final synthesis with markdown, tables, and structured reports * Editable agent cards with per-agent model selection * Reusable presets for telecom, code review, RCA, scrum, and more